Ανάπτυξη μεθόδων βαθιάς μάθησης για την ανίχνευση και ταξινόμηση μελανωμάτων
Loading...
Date
2026-09-25
Authors
Γιαλέρνιος, Παναγιώτης
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis investigates automated skin-lesion segmentation and multiclass classification,focusing on two controlled interventions: the use of generated lesion masks to condition the input and the depth of fine-tuning applied to pretrained models. A unified and reproducible experimental protocol was developed using fixed data splits, leakage checks, multiple random seeds, paired comparisons, measured noise floors, and external evaluation on a deduplicated ISIC 2019 cohort. In addition to conventional performance metrics, the study assessed calibration, uncertainty, out-of distribution detection, selective prediction, corruption robustness, learned representations, and attribution behavior.For segmentation, UNet++, Swin-Unet, and SapFormer remained statisticallyindistinguishable under the specified protocol. A soft-voting ensemble yielded a small positive advantage and was therefore used as the mask generator. In the classification benchmark, Swin-T and EfficientNet-B4 formed a statistically indistinguishable upper tier; Swin-T was selected as the common workhorse because of the thesis's transformer-based framing, rather than as an absolute numerical winner.The generated masks were technically adequate but did not provide a practically meaningfulclassification improvement over the raw input. In the final five-seed comparison, the mean change in macro-AUC was only +0.0024, and no corresponding benefit was detected on the external cohort. By contrast, moving from the overall M0 regime to M1 improved ranking metrics and substantially changed both the learned representations and the spatial concentration of attribution within the lesion.Increasing adaptation depth from M1 to M2 produced no measurable additional benefit across most predictive, representational, and explainability endpoints, indicating saturation. The same qualitative pattern was observed for ConvNeXt-T and ConvNeXt-ST-AFF, although these supplementary ladders were limited to one seed.Improved ranking did not imply uniformly improved reliability. M0 was better calibrated,whereas M1 performed better in ranking and robustness analyses. Moreover, under M1 the melanoma false-positive rate was approximately 2.87 times higher for images with heavy vignetting than for images without vignetting, while no fine-tuning-rung difference in this susceptibility was resolved.The AFF analyses further showed that the validity of representation and attribution methods is architecture dependent. Overall, the results favor controlled, architecture-appropriate, and adequately replicated adaptation choices over indiscriminate increases in model or training complexity.
Description
Keywords
Μελάνωμα, Βαθιά μάθηση, Ταξινόμηση δερματικών βλαβών, Τμηματοποίηση δερματικών βλαβών, Μετασχηματιστές όρασης, Προσαρμογή προεκπαιδευμένων μοντέλων, Εξηγήσιμη τεχνητή νοημοσύνη, Αξιοπιστία και ανθεκτικότητα μοντέλων